Dissecting psychiatric spectrum disorders by generative embedding.

Area of Science:

• Neuroscience
• Computational Psychiatry
• Machine Learning

Background:

• Psychiatric spectrum disorders are complex, often lacking clear biological markers for subgrouping.
• Current diagnostic methods may not fully capture underlying neurophysiological mechanisms.

Purpose of the Study:

• To investigate the feasibility of using generative embedding with dynamical system models to define subgroups in psychiatric disorders.
• To compare the efficacy of model-based neuroimaging analysis against conventional methods for clinical classification.

Main Methods:

• Re-analysis of fMRI data from schizophrenia patients and healthy controls during a working-memory task.
• Application of dynamic causal modeling (DCM) to infer neuronal circuit mechanisms.
• Utilized supervised (support vector machine) and unsupervised (Gaussian mixture modeling) learning techniques on DCM parameter estimates.

Main Results:

• DCM-based effective connectivity significantly outperformed functional connectivity and local activity for predicting diagnostic labels (78% vs. 62% vs. 55%).
• Unsupervised clustering identified two groups (patients and controls) with high accuracy (71%).
• Within patients, three neurophysiologically distinct subgroups were identified, correlating with clinical differences in negative symptom severity (PANSS).

Conclusions:

• Generative embedding using DCM provides a powerful, interpretable method for classifying psychiatric disorders.
• This approach can reveal clinically relevant neurophysiological subgroups within patient populations.
• The findings support the use of generative models for advancing computational approaches in clinical neuroscience.